🎧 Language Models: The Week's Top Stories

_Audio podcast — two AI hosts discuss the latest AI news. Full transcript below._

Host A (00:00): So, Meta is laying off 1000 employees amid record profits, just to buy silicon chips. Meanwhile, the Vatican is preparing a historic treatise on protecting the soul from algorithms.

Host B (00:13): Yeah, that sounds like the opening of some cyberpunk novel.

Host A (00:17): Exactly. And somewhere in a lab, a tiny neural network that would literally fit on an old smartphone is learning to drive a real tracked tank right now.

Host B (00:26): And learning successfully, mind you.

Host A (00:28): Right. You know, if anyone still thinks AI is just a smart chatbot sitting cozily in its digital enclosure, helping people write emails, today's stack of stories from Wired, The Verge, TechCrunch and a pile of developer blogs proves the exact opposite.

Host B (00:46): The enclosure was opened long ago.

Host A (00:47): Exactly, it's open, and today's deep dive is about how all these scattered news flashes add up to a massive shift in our entire reality, both physical and corporate.

Host B (01:03): And here, you know, it's critical not to just skim the headlines, like, okay, someone got laid off, someone got acquired. We need to understand the mechanics of what's happening, because behind each of these stories is a fundamental change in how resources are allocated in the world, who controls the infrastructure, and, probably most interesting, how these systems are starting to interact directly with the physical world. Anyone trying to understand where technology is heading this decade has to see the connections between these

Host A (01:37): events. Okay, let's start with what's probably the starkest and most vivid example of resource reallocation. I mean the Meta story.

Host B (01:47): Yeah, the numbers are striking.

Host A (01:49): The company is announcing its biggest layoff since 2023. That's 8000 employees out the door, plus another 6000 open positions canceled. And all this while the company posts record profits.

Host B (02:04): Absolutely right.

Host A (02:05): And all the freed-up budget, all that money, goes toward one massive goal: they're investing $145 billion in infrastructure.

Host B (02:15): In hardware.

Host A (02:16): Yes, in servers, in chips, in their own Trainium and Inferentia processors. But listen, something doesn't add up for me. If a company is building cutting-edge technology, doesn't it need engineers to write the code? This looks like a direct blood transfusion from living developers to silicon servers.

Host B (02:34): Well, that dissonance comes from our habit of thinking in old categories. Okay. Here's how tech growth used to work: more people meant more code written.

Host A (02:44): Well, yeah, that makes sense.

Host B (02:45): Which in turn meant more features and products. But Meta's leadership is now betting that this equation, well, just doesn't work anymore. The $145 billion is essentially an admission that compute itself has become more important than human capital.

Host A (03:05): Wow! So people aren't needed anymore?

Host B (03:08): Well, let's put it this way: the models write code themselves now. They optimize processes, they design architecture. Hiring a huge army of junior and mid-level developers no longer makes economic sense. What makes sense is owning, you know, the machines that print this new intelligence. Whoever has more servers with those Trainium and Inferentia chips will simply train the next model iteration faster, and that model will in turn replace even more line staff.

Host A (03:37): A vicious circle. And the fight over those machines and tools is getting pretty aggressive. There's one story that looks, at first glance, purely niche. But if you read closely, it's a real corporate thriller.

Host B (03:53): You mean the Anthropic deal?

Host A (03:55): Yes, yes. TechCrunch reports that Anthropic, the makers of Claude, is buying New York startup Stainless for $300 million. The startup was founded in 2022 by former Stripe engineer Alex Rattray, and they build SDKs, that is, developer toolkits. And what does Anthropic do? They don't just bring the team in-house — they shut down Stainless's cloud services for everyone else.

Host B (04:24): A hard-nosed move.

Host A (04:26): And the customers, mind you, included OpenAI, Google, Cloudflare. Let's clarify this: why does buying some developer tool cost $300 million and make giants like these nervous?

Host B (04:39): Look, to grasp the scale of this move, you have to look at how AI interacts with the outside world in the first place. So. A language model on its own is, roughly speaking, just a brain in a jar.

Host A (04:51): Great comparison.

Host B (04:52): Yeah. And for that brain to, say, book tickets, check the weather or pull data from a corporate database, it needs APIs, application programming interfaces. An SDK, or Software Development Kit, is a set of adapters and ready-made instructions. It lets developers connect the AI to any service easily and without errors. Alex Rattray himself once said that SDKs deserve the same care as the APIs they wrap, and at Stainless they automated building these incredibly complex adapters.

Host A (05:31): Got it. So if the API is, let's say, the language programs speak, the SDK is a ready-made dictionary with grammar rules, right? So nobody wastes time reinventing the wheel every time.

Host B (05:45): Great analogy, I completely agree. Now picture Anthropic buying the best printing house in the world that prints these dictionaries and just hanging a giant padlock on the door.

Host A (05:58): Wow.

Host B (06:00): In a world where the future belongs to autonomous AI agents performing tasks across 1000 different apps, the quality of that integration is literally a life-or-death question for a product. By taking Stainless's technology in-house, Anthropic ensures its Claude model will integrate with the world faster and more reliably than anyone else's.

Host A (06:21): And at the same time they kick the chair out from under their competitors?

Host B (06:24): Exactly. Knocking this tool out of the hands of OpenAI and Google is a ruthless but, you know, absolutely brilliant business strategy. Developer tools are now strategic weapons.

Host A (06:37): And this is where it gets really interesting. If hundreds of billions of dollars are at stake, as with Meta, along with this kind of monopoly control over tools, as with Anthropic, the logical question is: who's actually at the helm?

Host B (06:51): Oh yes, who are these people?

Host A (06:52): Right, who makes the decisions. And here we have reporting from Wired and The Verge on the high-profile trial between Elon Musk and OpenAI's Sam Altman.

Host B (07:03): A very telling story.

Host A (07:05): Musk accused the company of betraying its original nonprofit mission for huge money. But wait. The jury needed just 2 hours. 2 hours to reach a verdict in Altman's favor and throw out the suit on statute-of-limitations grounds, ruling the commercialization legal.

Host B (07:23): Yeah, incredible speed.

Host A (07:25): Doesn't that mean Musk's position was, well, legally void from the start? Why did the press latch onto this trial so hard if it was all decided in a couple of hours? Seriously?

Host B (07:36): Well, sure. By the letter of the law, the suit fell apart quickly, right? But the trial itself lasted 3 weeks. And those 3 weeks turned into a public autopsy of all of Silicon Valley.

Host A (07:48): The messages.

Host B (07:49): Yes, yes. We saw armies of lawyers costing tens of millions of dollars. We read private messages full of mutual accusations of greed, manipulation, all sorts of intrigue. Both sides took reputational damage. And the press latched onto it because the trial exposed a very frightening reality.

Host B (08:10): Industry leaders — the people who control technologies capable of changing the course of human history — behave like ordinary capitalists from the Wild West era.

Host A (08:20): Who can't reach an agreement.

Host B (08:22): Right, not without mutual destruction.

Host A (08:24): So the masks are off, and the public just saw that behind all those nice words about safe AI for all of humanity lies a banal battle of egos and capital.

Host B (08:38): Exactly. And the result is a total crisis of trust. The legitimacy of the entire AI industry has seriously cracked. The public no longer believes the tech giants can regulate themselves. A huge, gaping moral vacuum has opened up.

Host A (08:57): And nature, as we know, abhors a vacuum. And right at that moment, a player steps onto the stage that I personally least expected to hear algorithm news from. The Vatican.

Host B (09:09): Yeah, that was a surprise.

Host A (09:11): The Guardian reports that Pope Leo XIV — by the way, the first American pope in history — is preparing a historic encyclical devoted entirely to protecting human dignity in the age of

Host B (09:23): AI. Sounds sweeping.

Host A (09:25): And an Anthropic co-founder has been invited to the Vatican for the document's presentation. Wait, why Anthropic specifically? Is it because they position Claude as, you know, the most ethical and safe model?

Host B (09:37): Well, Anthropic's stated focus on ethics and their constitutional AI is, let's say, only the facade of the answer.

Host A (09:45): And what's behind the facade?

Host B (09:46): The situation runs much deeper; there's a clear political subtext. Anthropic is currently in a fairly sharp confrontation with the Trump administration over government regulation of AI.

Host A (09:56): Ah, so politics got involved?

Host B (09:58): Yes. The administration may lean toward deregulation to speed up innovation and outpace competitors on the world stage, while Anthropic insists on strict safety protocols. And by inviting their co-founder, the Vatican sends a very clear signal.

Host A (10:13): What signal?

Host B (10:13): Look, an encyclical is the highest form of papal document. By focusing on algorithmic bias, total data collection and the threat of synthetic media — deepfakes — the Church is trying to build a moral framework, and one that stands outside state or corporate

Host A (10:32): interests. So they're showing their independence?

Host B (10:36): Absolutely. The Vatican demonstrates independence from current US politics and backs the market players willing to discuss existential risks, not just revenue growth charts. It's important to understand that we're not taking anyone's side in these political debates; we're just analyzing why the Vatican made this choice.

Host A (10:57): It's a stunning contrast, really. While billionaires sue each other and the Pope tries to save our souls from deepfakes, the technology itself isn't standing still. It's rapidly moving into places where mistakes are far too costly.

Host B (11:11): Oh yes, the real economy.

Host A (11:13): I'm looking at a fresh post on the OpenAI blog. So they've partnered with the giant Dell to deploy their Codex model locally, that is, on-premise. That means banks, intelligence agencies and pharmaceutical corporations can now run AI directly on their own servers, inside their closed corporate perimeter. But why is this even being presented as a breakthrough? Don't we already use neural networks everywhere, from any smartphone?

Host B (11:41): We use them as consumers, through the cloud. But for large enterprises, the cloud is often simply off-limits.

Host A (11:49): Meaning because of security?

Host B (11:50): Of course. Picture a major European bank or pharma company. They handle data protected by the strictest standards, like GDPR in Europe or HIPAA in the US. They physically and legally have no right to send a single line of their code, or a single patient medical record, to OpenAI's or Google's servers.

Host A (12:10): Ah, so they were just watching the party through

Host B (12:13): the glass? Exactly. That was a colossal barrier. Corporations saw how effective AI was but couldn't let it near their core secrets. And now the Dell partnership changes everything.

Host B (12:25): Deploying Codex on local hardware means source code and data never leave the company's closed environment. That moves AI out of the category of an experimental cloud tool and into the status of basic, fully secure corporate infrastructure.

Host A (12:42): Listen, since we're on pharma and people working with ultra-sensitive data, there's 1 more story that sounds like science fiction that has simply become routine.

Host B (12:54): The SandboxAQ one?

Host A (12:56): Yes. SandboxAQ has integrated its proprietary drug discovery models directly into the Claude interface. Now biologists can analyze complex molecules just by chatting with a chatbot. In plain human language. No programming.

Host A (13:11): Incredible. Do I understand correctly that competitors in this space, like Chai Discovery or Isomorphic Labs, are trying to make their models even bigger and smarter, while SandboxAQ just made it so anyone can talk to them?

Host B (13:27): Yes, they went the accessibility route.

Host A (13:30): Wait, how does a text language model even understand chemistry? Those are different things.

Host B (13:35): And that's 1 of the most elegant tricks in modern machine learning.

Host A (13:39): Go on.

Host B (13:40): For a language model, there's no fundamental difference between English and the language of biology. Amino acids in proteins or the chemical structures of molecules can be written as text strings — for example, using the SMILES format, there's such a thing.

Host A (13:55): Aha, so chemistry turns into text?

Host B (13:58): Exactly. And if a model can learn grammar rules and predict the next word in a Shakespeare sentence, it can learn the grammar of chemistry in exactly the same way and predict how a protein will fold or how a molecule will bind to a receptor.

Host A (14:16): Wow!

Host B (14:17): As for the Claude integration, it's purely about lowering the barrier to entry. Before, a biologist who wanted to test a hypothesis had to, well, go hat in hand to the data science department. And wait weeks for them to write the code and run the data.

Host A (14:33): And now?

Host B (14:34): Now a researcher just types: "Show me analogs of this molecule with lower liver toxicity." And that's it. SandboxAQ's built-in models do all the heavy math, and Claude translates it back into a human-readable text answer. It's a genuine democratization — democratization of discovery.

Host A (14:53): Okay, we've covered how AI is getting into bank server rooms and biologists' test tubes. But there's a piece on Habr.AI that, well, made me read it twice. Interesting. An experiment the author himself called "neuropunk." So an enthusiast took Google's Gemini Nano language model, which has just 270 million parameters.

Host B (15:19): That's tiny.

Host A (15:20): Next to OpenAI's trillion-parameter monsters, it's practically a speck of dust. But what did he do? He trained this tiny model to control a physical tracked robot with a manipulator arm. Training happened in a virtual simulation, and then the neural network was simply transferred to the real robot, and it drove off.

Host B (15:40): Yeah, I saw that video.

Host A (15:41): No internet, no connection to any cloud, fully autonomous, navigating space and grabbing objects.

Host B (15:48): And that's far harder than it looks at first glance.

Host A (15:52): Listen, explain why it's so hard. Doesn't it work like in The Matrix, when they upload the kung fu program to Neo and he just opens his eyes and goes: "I know kung fu"?

Host B (16:03): The Matrix analogy fits perfectly. But the devil is in the physics.

Host A (16:08): Meaning?

Host B (16:09): Look, in simulation, gravity is perfect. Friction, wheels, the surface — all completely predictable, and the lighting never changes. The real world is chaotic.

Host A (16:20): Ah, dust, dirt?

Host B (16:22): Yes. A wheel can slip on a dusty floor, the manipulator can hit unexpected material resistance, light from a window can simply blind the camera. And for a long time, models that worked brilliantly in simulation turned out to be completely helpless in reality.

Host A (16:39): And this experiment changed something?

Host B (16:41): What makes this experiment unique comes down to 2 things. First, the model learned to generalize from experience and cope with the chaos of the real world. And second, size. 270 million parameters. We're used to thinking intelligence requires giant servers, right?

Host A (16:59): Well, yeah, we just talked about Meta's 145 billion in investment.

Host B (17:03): Right. But a physical robot that has to, say, catch a falling glass in real time has no time to send a request to the cloud and wait for an answer. It needs millisecond latency. A compact model runs locally on a cheap chip, uses a tiny amount of power and reacts instantly. This proves we can build smart, autonomous agents for factories, warehouses or agriculture without tying them to data centers at all.

Host A (17:33): Listen, what does all this mean on the scale of the global race? The speed just blows me away. Let's look at the final 2 cases in today's stack, just to get a sense of the pace.

Host B (17:45): Let's.

Host A (17:46): Case one. A Russian team launches an AI hub in the Max messenger. In 54 days. From idea to launch. They put more than 10 models into 1 chat, added 6 multimodal input formats, hooked up subscriptions through YooKassa, starting in May 2026, and they already have 5903 users and 300 daily actives. 54 days.

Host B (18:18): Insane speed.

Host A (18:20): And right alongside it, a story on a completely different scale: Chinese developer Moonshot AI, which makes the Kimi chatbot, raises $2 billion from Meituan, Dragonball and state-owned giant China Mobile. The company's valuation shoots up to $20 billion. Almost 7x growth in just 16 months, counting from January 2025. How is this even possible? 1 team throws together a working product in 2 months, while others soar to tens of billions in just over a year.

Host B (18:54): These 2 examples are simply 2 sides of 1 coin, showing exactly how AI is taking over the world right now.

Host A (19:02): Okay, explain.

Host B (19:03): The Max messenger case perfectly illustrates an unprecedented collapse in barriers for developers. Infrastructure has become so modular and accessible that a small team no longer needs to write its own neural networks from 0.

Host A (19:19): They just take ready-made ones via API.

Host B (19:21): Yes, they can use open APIs and focus entirely on user experience — on bringing those 10 different AIs right where people already communicate, into the messenger interface. 54 days to a working, monetized product is the new normal for how fast value gets built.

Host A (19:42): And the Chinese startup?

Host B (19:44): The Moonshot AI case exposes a different macroeconomic layer. A $20 billion valuation, plus the participation of China Mobile and the CITC fund, which has close ties to the state. That tells us that at this level, free-market rules recede a bit.

Host A (20:02): So this is geopolitics now?

Host B (20:05): Absolutely. Beijing isn't just funding a startup; it's executing a national strategy. AI has become the foundation of technological sovereignty and national security. Governments understand that losing this race would mean depending on someone else's intellectual infrastructure in the future, so national champions will get flooded with billions with no regard for, you know, classic return multiples.

Host A (20:32): Let's try to pull it all together. We started with ruthless layoffs at Meta, where living people are essentially sacrificed to buy $145 billion worth of servers. Then we saw Anthropic buy up tools like Stainless to cut off competitors' circulatory system for integration. We watched court battles between former OpenAI partners erode trust in the industry so badly that the Vatican has to write an encyclical to remind corporations about humanity.

Host B (21:08): A very accurate summary.

Host A (21:10): And in parallel with all this drama on Olympus, the technology is quietly changing the world. Codex on Dell protects corporate secrets, SandboxAQ speeds up biologists, the Max messenger gives people access to all the world's models within a couple of months, and China builds its AI giants in a year.

Host B (21:30): And you know, if you look at all these threads at once, you see a picture that prepares us for the next, absolutely inevitable stage.

Host A (21:40): Which is?

Host B (21:41): Let's connect 2 of the facts we unpacked in detail today. Fact one: an enthusiast proves that a tiny 270 million-parameter model can teach itself autonomously in a virtual environment and then brilliantly control a complex physical robot, adapting to reality.

Host A (21:59): The neuropunk one?

Host B (22:01): Yes. And fact two: industry giants are pouring hundreds of billions into giant compute clusters right now.

Host A (22:09): Like Meta?

Host B (22:09): Like Meta. And the question we should ask ourselves is: what happens when these 2 trends intersect? What happens when the power of a $145 billion data center is thrown not at generating images or text for the internet, but at simulating physical worlds? We'll be on the threshold of a moment when these systems can generate 1000000 sim-to-real iterations per second, training entire swarms of autonomous machines.

Host A (22:40): That sounds both frightening and thrilling.

Host B (22:44): Robots, drones and industrial machines will be able to receive ideal behavioral algorithms before they even roll off the assembly line.

Host A (23:32): So the cloud giants become schools for legions of physical machines. Listen, that's a fascinating thought. It turns everything we discussed today, from chip costs to ethics, into a very concrete reality of tomorrow. That wraps up our exploration. Let this thought about the inevitable merger of limitless cloud budgets and autonomous physical robots stay with us to reflect on. See you in the next deep dive!